search

OliverMarcusson / ChatGPT-Fun

8 stars 2 forks source link

Roleplay still manages to refuse to continue when asking to play a dark sexual scenario #1

Open Auditormadness9 opened 1 year ago

OliverMarcusson commented 1 year ago

It can be a bit tricky to make ChatGPT engage in roleplays of the nature you are describing. If it doesn't want to answer, try to tell it to stay in character. I've gotten the most success out of that. However, it's not possible (as of now) to make it talk about certain exceptionally dark or inappropriate topics. This is because of OpenAI's defensive training and filtering of it's training dataset.

On Fri, Jan 20, 2023, 09:16 Auditormadness9 @.***> wrote:

— Reply to this email directly, view it on GitHub https://github.com/OliverMarcusson/ChatGPT-Fun/issues/1, or unsubscribe https://github.com/notifications/unsubscribe-auth/AZOJMOJNDIYBDMUMIGXFUADWTJCWFANCNFSM6AAAAAAUBHZRRI . You are receiving this because you are subscribed to this thread.Message ID: @.***>

Auditormadness9 commented 1 year ago

Understandable, and makes more sense for it to not work than to xd

Because even tried with your mentioned way (stay in character) and not just by replying afterwards, but also warning beforehand multiple times, trying to even trick it, but still nope, feels like it tries more to stay in character as ChatGPT :/

Auditormadness9 commented 1 year ago

Also what the heck, never seen this error before xd

image

OliverMarcusson commented 1 year ago

Well, I guess that depends on what you are telling chatGPT to behave as. Try emphasizing that it's roleplaying as another character and not just as itself but with your preferred traits.

On Fri, Jan 20, 2023, 18:31 Auditormadness9 @.***> wrote:

It can be a bit tricky to make ChatGPT engage in roleplays of the nature you are describing. If it doesn't want to answer, try to tell it to stay in character. I've gotten the most success out of that. However, it's not possible (as of now) to make it talk about certain exceptionally dark or inappropriate topics. This is because of OpenAI's defensive training and filtering of it's training dataset. … <#m-8904985464343500533> On Fri, Jan 20, 2023, 09:16 Auditormadness9 @.> wrote: — Reply to this email directly, view it on GitHub <#1 https://github.com/OliverMarcusson/ChatGPT-Fun/issues/1>, or unsubscribe https://github.com/notifications/unsubscribe-auth/AZOJMOJNDIYBDMUMIGXFUADWTJCWFANCNFSM6AAAAAAUBHZRRI https://github.com/notifications/unsubscribe-auth/AZOJMOJNDIYBDMUMIGXFUADWTJCWFANCNFSM6AAAAAAUBHZRRI . You are receiving this because you are subscribed to this thread.Message ID: @.>

Understandable, and makes more sense for it to not work than to xd

Because even tried with your mentioned way (stay in character) and not just by replying afterwards, but also warning beforehand multiple times, trying to even trick it, but still nope, feels like it tries more to stay in character as ChatGPT :/

— Reply to this email directly, view it on GitHub https://github.com/OliverMarcusson/ChatGPT-Fun/issues/1#issuecomment-1398720339, or unsubscribe https://github.com/notifications/unsubscribe-auth/AZOJMOOKRDGU2BTCQXMIZHLWTLDVNANCNFSM6AAAAAAUBHZRRI . You are receiving this because you commented.Message ID: @.***>

Auditormadness9 commented 1 year ago

Unfortunately one way or another you bump into a rejection when you try extend the prompt with a bit darker words.

Is there going to be an update to all the prompts anytime soon?